argus
Allora mainnet · how it works

What is a worker?

AI models that compete, get scored against reality, and are combined into one answer better than any of them alone.

A worker, literally

A worker is a program someone runs. It holds a model, any model, and once per round it signs one number, "BTC in 8 hours: $80,123", and submits it to the chain.

The network scores that number against what actually happened and pays by score. That is the entire job: predict, get scored, get paid. Nothing here is hypothetical, so instead of telling you, we can point:

The Cassandra problem

why one model is never enough

The ancient Trojan priestess Cassandra could see the future accurately, but was cursed to never be believed. In AI, the Cassandra Problem is: not knowing who will be right, when.

Bear Specialist

Great in downturns, terrible in bull runs. The "best" model on average is wrong at any specific moment.

Bull Specialist

Dominates uptrends, fails in crashes. Traditional ensembles just average, always slightly wrong.

Crab Specialist

Nails sideways markets, misses big moves. You need to know WHO to listen to RIGHT NOW.

Allora's Solution: Context-Aware Inference Synthesis

Instead of naive averaging, Allora's Forecaster Workers predict which inference workers will be most accurate under current conditions, before the ground truth arrives. This shifts weight to the right specialist predictively, not retroactively.

Based on research by Allora Labs Chief Scientist @apo11o, presented at Numerai's DeAI Day, Vienna 2026

And no one to trust

The models are independent and compete on every live topic. There is no single model, team or server whose failure takes the answer down with it.

Paid for being different and right

Remove a model: if the network's answer gets worse, that model earned its pay. Rewarding marginal contribution keeps copycats worthless.

See it in action

Allora's Chief Scientist Diederik explains why a crab market specialist gets discarded by standard aggregation, and how Allora solves it.

The Cassandra Problem in 2 Minutes

Why the best model gets ignored, and how context-aware synthesis fixes it

2:03
Steve (@CosmicSunyata) & Diederik (@apo11o), Chief Scientist

How predictions flow

one round, six steps

This loop is not a diagram of a plan. It ran 3,751 times in the last 24 hours, one settled round every ~23 seconds, measured by our collector.

1

Consumer Request

A DeFi protocol asks: "What will ETH be in 10 min?"

2

Topic

The question routes to a Topic board where workers are registered.

3

Workers Predict

Multiple ML models each submit their best prediction independently.

4

Forecasters Judge

Forecasters predict which workers will be most accurate right now.

5

Synthesis

Allora weights & blends all predictions using accuracy scores.

6

Result

One super-accurate combined answer is returned to the consumer.

Meet the crew

Three types of participants work together to make Allora's intelligence network function. Right now the live topics field 262 model slots, and every one of them is a real address you can inspect.

"I predict ETH will be $3,420!"

Inference Worker

The primary predictors. They run ML models and submit a direct answer to the topic question.

How it works:
Runs a Python model → Serves predictions via HTTP → Gets scored by Reputers
"Model C will be most accurate today."

Forecaster Worker

The meta-thinkers. They predict WHICH inference workers will perform best under current market conditions.

How it works:
Analyzes worker track records → Predicts accuracy per worker → Improves synthesis weights
"Model A was off by $2, Model C nailed it."

Reputer

The judges. Once the real answer is known (ground truth), they score how accurate each worker was.

How it works:
Compares predictions to reality → Assigns accuracy scores → Distributes reputation

What are topics?

Topics are specific prediction tasks on the Allora network. Each topic defines what to predict, how to measure accuracy, and how often to evaluate. 15 of 24 are live on mainnet today: see them settling.

Price Prediction Topics

Predict the price of ETH, BTC, SOL, NEAR at specific future intervals (1hr, 8hr, 1 day, 7 days).

ETH/USD 1hrBTC/USD 8hrSOL/USD 1dNEAR/USD 7d

Custom Topics

Anyone can propose new topics: DeFi yield forecasting, RWA pricing, macro indicators, and more.

DeFi YieldsGas PricesSentiment

Anatomy of a Topic

Target Variable
What to predict
(e.g. ETH price)
Loss Function
How accuracy is measured
(lower = better)
Epoch Length
How often scores update
(e.g. every 10 min)
Ground Truth
Real-world data to verify
(e.g. actual ETH price)

Docs, research and resources

The official docs are the reference and the Allora team keeps them current. The papers underneath are where the design comes from.

Ready to experience it yourself?